Files
MeanField/experiments/PRECONDITIONER_IMPROVEMENT_2026-09-18.md

148 lines
8.4 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Preconditioner improvement: matched frozen Newton systems
This experiment changes preconditioning only. The complete Newton Jacobian,
normalization, physical residual, mesh, and linear acceptance gate are held fixed.
The date in branch and artifact names is an experiment identifier; provenance
files contain the actual UTC execution times.
## Reproduction and scope
Integration branch: `research/preconditioner-benchmarks-20260918`, based on
`10adf7e`. Hypotheses are retained separately on
`research/gravity-consistent-schur-20260918`,
`research/preconditioner-cached-cross-actions-20260918`, and
`research/coupled-shape-coarse-20260918`.
The experiment workspace is
`/Users/tboudreaux/.codex/.chatgpt-projects/g-p-6a6b3f2194ec8191bebcd59075897f1d`.
Evidence is under `diagnostics/preconditioner_improvement_2026-09-18` there.
Each campaign retains its clean source commit, build configuration, copied
executable and hash, mesh hash, exact commands, input/output hashes, and logs.
Runs are serial, without overlapping builds or FEM jobs during timing.
Frozen comparisons start FGMRES from zero on the same saved state and right-hand
side. Each role constructs a fresh inverse, including its own specification
border responses. The restart length is 160, the initial cap is 200 iterations,
and the required relative full-J true defect is 0.03. The independently applied
full Jacobian decides acceptance; the Krylov recurrence flag alone does not.
All accepted/trial state buffers, normalization factors, preparation generation,
and a deterministic full-J probe must remain unchanged.
Costs include inverse construction, normalized binding, Krylov setup, Krylov
solve, and full-J verification. Extra diagnostic applications and artifact I/O
are recorded separately. A speedup to convergence is reported only if both
methods meet the same true gate. Less work spent on a failed solve is not a
time-to-convergence improvement. A frozen linear pass is not nonlinear or
physical convergence.
## Existing two-way coupling: negative controls
`screen_v1` uses source `018fc1c`, before the cached cross-action optimization.
| Saved state | Inverse | Iterations | Full-J relative defect | Complete verified cost (s) |
|---|---|---:|---:|---:|
| n=1.5, J=0 plateau | independent | 200 | 0.0590170471860 | 87.9302 |
| same | stellar LDU | 200 | 0.0600668598825 | 599.5473 |
| n=1, J=0.1 plateau | independent | 200 | 0.0826534298794 | 86.9409 |
| same | stellar LDU | 200 | 0.0720550396371 | 594.2053 |
All four miss the 0.03 gate. Stellar LDU spends 6.82–6.83 times the total cost;
it slightly worsens the n=1.5 result and reduces the rotating defect by 12.82%.
Mean inverse application costs rise from approximately 0.19 s to 2.7 s, while
full-J actions remain approximately 0.23 s. The uncached gravity-to-material
force column explains the application-cost anomaly.
Poisson contributes 95.53% of the squared n=1.5 baseline defect and 98.28% with
stellar LDU. For the rotating state those fractions are 85.00% and 98.10%; LDU
reduces the gradient-row norm by 94.16% but the Poisson norm by only 6.35%.
These are global normalized block diagnostics, not proof of an exclusively
exterior error. No exterior weighting is changed.
The independent artifact audit passes 65 provenance checks and 93 checks per
case. Both saved states and the normalization data exactly match their producer
artifacts. The two independent baselines reproduce the earlier failed solves.
## Candidate mechanisms
1. **Prepared cross action.** Reuse the complete prepared gravity force with
zero density and displacement directions to apply its gravity-gradient
column. This is the same mathematical coupling with cached geometry; it does
not strengthen the inverse. See `PRECONDITIONER_CACHED_CROSS_ACTION_2026-09-18.md`.
2. **Consistent gravity Schur refinement.** Keep the existing AMG action, but
apply a fixed number of corrections against `B R B^T`, where `R` is the same
mass inverse used in flux recovery. Fixed work preserves linear stationarity
for stationary child inverses. Damping does not guarantee contraction; it
must earn its additional cost. See `GRAVITY_CONSISTENT_SCHUR_2026-09-18.md`.
3. **Experimental coupled shape coarse inverse.** Use nine fixed geometric
modes, complete their material/gravity responses, and construct a checked
low-rank correction to the normalized inverse. Modes do not depend on the
Newton right-hand side or previous Krylov results. A rank/overlap guard
explicitly falls back to the underlying inverse when the correction cannot
be trusted. Coarse setup and storage are counted, and actual FEM subspace
identities are independently checked before solving.
## First candidate screen
`candidates_v1` uses source `dc8d4b0` and the same n=1.5, J=0 plateau state.
Every fresh independent control reproduces relative defect 0.059017047186 at
200 iterations. Complete control costs remain between 87.53 and 88.25 s.
| Candidate | Iterations | Full-J relative defect | Gate | Complete verified cost (s) |
|---|---:|---:|---|---:|
| Cached stellar LDU | 200 | 0.0600668598779 | fail | 117.4845 |
| Consistent gravity Schur, k=1, omega=0.5 | 132 | 0.0292091549685 | pass | 72.3981 |
| Cached stellar LDU + consistent gravity Schur | 121 | 0.0285619160766 | pass | 84.7104 |
| Nine-mode coupled shape coarse correction | 200 | 0.0668657989082 | fail | 97.0564 |
The cached LDU is 5.1032 times faster than its old execution path at the same
iteration cap, while mean inverse application cost drops 8.0425 times. This is
an execution improvement, not a convergence improvement. Saved inputs and the
full-J probe agree exactly; the final normalized correction changes by only
2.7051e-11 relatively. Both LDU solves fail, and the cached version remains
34.2% more expensive than the independent control on this state.
Gravity refinement is the first candidate to cross the original gate. It lowers
the Poisson-row defect by 55.4% and the gradient-row defect by 15.0%. The squared
Poisson share drops from 95.53% to 77.59%; surface and hydrostatic defects grow,
so the improvement is judged by the verified global norm. Work changes from
400 to 396 mass-inverse applications, 200 to 264 AMG Schur applications, and
200 to 132 outer inverse applications. The default has not yet passed the same
gate, so dividing its failed 88 s attempt by 72.4 s is not a convergence speedup.
The combined LDU saves another eleven iterations but is 17% slower overall.
The coarse candidate has all nine modes active, minimum overlap lower bound
0.79465, and no fallback. Its extra construction costs 8.6012 s, including mode
completion and eighteen setup Jacobian applications; it stores 28,491,680 bytes
in persistent vectors. The actual FEM subspace check passes. Nevertheless, its
200-iteration defect worsens by 13.3%. This is a negative numerical result for
these modes on this state, not evidence about every possible coarse space or
rapidly rotating model. The prototype remains confined to experiments.
## Validation status
The integrated library and experiment driver compile with the repository's
Release/ThinLTO configuration. Source dc8d4b0 passes 20 focused serial test cases
(441,978 assertions), including FEM equivalence, geometry/discretization refresh,
fixed-work algebra, and coarse inverse tests. Distributed unit and coarse tests
pass on two ranks (13 cases and 213 assertions per rank).
The broad real-FEM MPI run aborts because the existing METIS partition gives one
rank no compactified vacuum element, which the mapped mass constructor requires.
The failing log is retained; this campaign does not claim distributed FEM Schur
validation. See the gravity refinement note for the exact scope and evidence.
The evidence checker passes 22 lightweight tests. Independent reviews of the
cached LDU and gravity-only cases validate provenance, state/operator invariants,
true gates, and child-work accounting. Reviews are stored outside immutable
campaign directories.
Source fcb7833 adds owning experiment-only Newton recipes through the existing
ADL preparation interface. They use the same concrete prepared inverses as the
frozen benchmark and the existing context refresh/border lifecycle. The driver
supports default, gravity-Schur, and stellar-LDU recipes for ordinary Newton,
positivity, curvature, and positivity-curvature runs. This adds no public
physics-facing solver API. The updated driver builds successfully at87569c0.
Equal-accuracy hard-state confirmation and Newton/virial regressions are pending.
The default production recipe remains unchanged during validation.